Indoor Scene Recognition And Understanding Based On Visual Perception

Indoor scene recognition and understanding is a vital technology of the intelligent information processing. The relative mature technology has been using in a number of different fields, e.g. security and protection including human detection, object tracking, face recognition, internet information processing including image and content retrieval, robot including object seeking, scene understanding, barrier detection. Its utility and convenience for work and life attracts reaserchers all over the world.Indoor scene recognition and object detection based on deformable part-based model(DPM) are focused in this thesis. The main contents are as follows:1. Starting from Histogram of Oriented Gradients, we study DPM’s composing, matching and scoring at multiple scales, as well as the model’s training algorithm, which is be implemented by training a latent support vector machine using coordinate descent method, combining a hard negative mining to reduce negative examples. Such a model detects object from different parts at different scales synergistically, which provides strong robustness to variation of object’s pose and part. Meanwhile, DPM is combined with feature context by radial basis coding, giving the model more context information. Experiment results show that the model outperforms other detection algorithm, and that feature context increases the performance further.2. We utilize DPM to get a scene feature and overcome the disadvantage that the region of interest relies on manual annotations. By adjusting DPM’s training, matching and parameters, DPM is used to implement scene recognition. Along with specific experiments, the model’s adjusting and the parameters’ influences are discussed in detail.3. In consideration of that DPM holds global scene insufficiently, the feature channel-based visual perception scene gist is introduced. We study the gist extraction of intensity, color, orientation and the fushion stratedy based on the softmax for scene recognition. Experiment results demonstrate that the fushion improves the performance and the proposed method outperforms the other ones.